XGBoost: The Split Gain
Intermediate
The problem, from first principles
XGBoost chooses splits by how much they reduce the regularized objective. The gain compares the score of the two children against the score of the parent, then subtracts a penalty for adding a leaf.
From theory to code
Implement xgb_split_gain(GL, HL, GR, HR, lam, gamma), returning the gain of a split minus gamma.
Constraints
- The parent sums are the children sums added together.
Hints
Hint 1
Compute each child's score G^2 / (H + lam), subtract the parent's score, halve the result, then subtract gamma.
Keywords
research-papersclassical-mlboostingxgboost
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.